Abstract

Information on paper documents is accessible to those who are around where they are stored. Such important documents are not available in softcopy for search engines. A lot of Yoruba historical documents were handwritten before now. Automated Character recognition is relatively common for English language but not for Yoruba language due to the diacritics which lead to tonal difference in Yoruba text. This research aims to design and implement a Word-level Yoruba Handwritten Character Recognition system using modified Vision Transformer for character recognition. The Yoruba handwritten images consist of handwritten Yoruba alphabets from a newly created database locally acquired. The newly created dataset contains 6434 images with 40 classes. 6414 images were used as training dataset while 20 word images were used as testing dataset. Each image was resized to 72x72 pixels and then converted to grayscale. Noise was removed from each grayscale image using Otsu algorithm and bounding boxes for word image segmentation. Vision transformer was used for recognition and 21,692,904 trainable parameters were obtained from initial empirical experiments. Average segmentation, character-level recognition accuracy and testing time of 79.7%, 72.1% and 0.9s respectively were obtained. The developed system gave a better performance when evaluated with recognition accuracy and testing time which outperformed SVM and CNN classifiers when compared with existing studies. The developed system accommodates more Yoruba vocabulary and successfully recognized common Yoruba words with acceptable testing time. The developed system would be useful in archiving and digitizing Yoruba historical handwritten documents and teaching and learning in Yoruba language